Your LLM Is Not a Colleague

Software engineer and blogger Dan H. argues that calling an LLM a colleague or teammate creates a dangerous excuse for shifting blame, noting that he is on pace to merge around 650 AI-assisted pull requests this year without ever saying 'the model did that.' He distinguishes harmless anthropomorphism from blame transfer, citing Cognition's Devin marketed as 'the first AI software engineer' and Coworker.ai as examples of industry marketing that encourages personification, and warns that this framing costs engineers credibility by signaling they shipped code they didn't understand.

Your LLM Is Not a Colleague The industry spent two years teaching you to call your agent a teammate. Now watch what happens the first time something breaks. Your LLM is not a colleague, and the moment you start talking about it like one, you are setting up an excuse you will eventually use. I’m on pace to merge around 650 pull requests this year, almost all of them AI-assisted. I have never once said “the model did that.” I keep hearing it, though. In standups, in Slack, in code review. There are two versions of it, and people keep conflating them, so let’s separate them first. The version that’s fine “Claude and I dug into this.” “I paired with the agent on the migration.” Harmless. We have always talked about tools this way. The compiler complains. The GPS took me down a dirt road. The build is being difficult today. My car doesn’t want to start. Nobody has ever filed a grievance about the compiler’s feelings. Anthropomorphizing tools is just how language works. If your objection is grammatical, you’re going to lose that argument, and you should. It’s the wrong argument. The version that isn’t “Oh, my LLM did that, what the hell.” That is not a figure of speech. It’s a transfer of blame, aimed at something that cannot receive it. Try the calculator test. Nobody has ever said “me and my calculator got the wrong answer.” The phrasing isn’t what stops them. It just wouldn’t work. Everyone in the room knows the calculator is yours, the inputs were yours, the answer is yours. Say it and you don’t sound blameless. You sound like someone who doesn’t understand what a calculator is. The LLM is the first tool in our profession plausible enough as an agent that the deflection actually lands. You say “the model decided to refactor that” and people nod, because it does kind of decide things. That’s new. It’s also the entire problem. You didn’t invent this. It was sold to you. I don’t blame the people saying it, and this is why. Nobody arrived at “my agent did that” on their own. The whole category has been marketed in the language of personhood since day one. Cognition launched Devin as “the first AI software engineer” https://cognition.com/blog/introducing-devin , not “a coding tool.” There’s a company called Coworker https://coworker.ai . Not a metaphor, that’s the name on the building. The sell side is completely open about it. Jason Lemkin, writing at SaaStr about what AI products should aim for in 2026 https://www.saastr.com/your-critical-product-goal-for-2026-your-customers-truly-honestly-see-your-ai-agent-as-part-of-the-team-for-real : The customer gives it a name. I’m not kidding about this. When customers start calling your AI agent “our Alex” or “our Sarah” instead of “the chatbot”, that’s when you know you’ve crossed the threshold. And the success metric, stated plainly: Would they actually miss it? Would they feel like they lost a teammate? If the answer is no, you have a problem. A stated product goal, not an accident of copywriting. Make the user feel like the software is a person, because people who feel that way don’t churn. Good business logic. I understand it. But it ships with a side effect nobody puts on the box: a teammate is blameable. Once the tool has a name and a personality and a seat in your standup, there is somewhere for responsibility to go when things break. And responsibility, given somewhere to go, will go there. What it actually costs you The industry’s favorite story is that AI’s victim is the junior engineer. I’ve argued before /blog/ai-wont-kill-juniors-it-will-expose/ that I don’t buy it. The person this costs is you, the engineer using the phrase. And it costs you in a way you probably haven’t noticed. When you say “the model did that,” you think you’re deflecting. What you’re actually announcing, out loud, to your team, is that you shipped something you didn’t understand. That’s the literal content of the sentence. There is no reading of it where you come out looking like you were in control. Worse, it kills the conversation. I can review a person’s reasoning. I can ask “why did you take this approach” and get an answer, and we both get smarter. I cannot review a ghost. “The LLM decided to” is a dead end, a shrug with a technical accent. The most useful thing about a code review is finding out what the author was thinking, and you’ve just told me there wasn’t an author. The delegation still has to be verified /blog/the-hidden-corruption-tax-of-ai-delegation/ by somebody, and verification requires an author to interrogate. Say the true thing instead The constructive version is not “speak more precisely about your tools.” Nobody’s going to do that, and I don’t care if they do. If you shipped something you didn’t fully review, say so. “I didn’t read this closely enough” is a better sentence than “the LLM did it.” It’s shorter, it’s honest, and unlike the ghost story it points at something a team can fix: your review process, your test suite, your prompt, your context. Every one of those is actionable. “Claude did it” is actionable by nobody. None of this is an argument for using agents less. I use them harder than anyone on my team, and I intend to keep going. Run twenty in parallel if that’s your workflow. Let them write the code, let them review each other’s code. Just don’t let them sign it. The commit has your name on it, and that was always the deal. Related posts How to Be a Great Software Engineer in 2026 /blog/how-to-be-a-great-software-engineer-in-2026/ The framework hasn't changed. The weight of each skill has. Read more → /blog/how-to-be-a-great-software-engineer-in-2026/ The Flow Is Gone /blog/the-flow-is-gone/ I used to hold entire systems in my head. Now I hold seven terminals. The trade was worth it, but something real got lost. Read more → /blog/the-flow-is-gone/